Kimi K2.5: Chinese 'Smarty' Beat OpenAI and Went After Dollars

Imagine stepping into the ring against a heavyweight with a hundred times less money in your pocket and a team of a couple hundred people against a multi-thousand-strong army. It sounds like a script for a sports drama, but that's exactly what Chinese startup Moonshot AI is pulling off with its Kimi model. While the industry held its breath waiting for OpenAI's next move, the folks at Kimi released K2.

5 update and suddenly realized that the Western market is willing to pay them more than their home Chinese market. This is a rare case when a local project from China doesn't just copy Western counterparts but starts dictating its own rules of the game on the global stage. The story of Kimi is a journey from "just another chatbot" to a sophisticated ecosystem of intelligent agents.

If the first version K1.5 focused on working with long contexts, the fresh K2.5 bets on "teamwork."

Creators implemented the concept of Agent Swarm, where up to 100 virtual agents can simultaneously execute up to 1500 steps to solve a single task. This is no longer linear dialogue but a real brainstorm inside a neural network. In tests like HLE (Human Last Exam) and SWE-Bench, the model shows results that make developers of Claude 3.

5 and GPT-4o nervous. Essentially, Kimi K2.5 currently occupies an honorable third place in the global OpenRouter ranking, which for a startup with such a budget looks almost like a miracle.

Why is this important right now? We've hit a ceiling with classical scaling. As Kimi founder Yang Zhilin correctly noted, internet data for training is running out faster than computational power is growing.

There's only one way out—scale not just the training but also the time the model spends on "thinking" (test-time scaling). Kimi chose the path of creating a swarm of agents that can simultaneously collect data, write code, and assemble presentations. This transforms AI from a secretary into a full-fledged production department.

The fact that the company's overseas revenue quadrupled after the release confirms: business is tired of empty promises and wants tools that really automate complex chains of actions. Moonshot AI's strategy resembles a hybrid of Anthropic and Manus. From the former they took the focus on fundamental research and algorithm quality; from the latter—ambitions to create universal agents.

Yet Kimi remains surprisingly compact: about 300 people on the payroll. This allows them to be flexible and implement innovations like the Muon optimizer or Linear Attention faster than corporate giants manage to agree on a coffee budget. They consciously limit their business boundaries, focusing on logic, coding, and deep research without trying to do everything.

For the global market, this is a powerful signal. Silicon Valley's monopoly on "smart" models has finally crumbled. When a company with 1% of Google's resources delivers comparable or even better results, the rules of the game change.

Now the question is not who has more H100 video cards, but whose algorithms handle available computing power more elegantly. Kimi has already started opening its weights and tools to developers, building around itself a loyal community that values efficiency above marketing noise. The bottom line: Kimi proved that the strategy of "smart scaling" through agents works better than simply scaling up parameters.

Can they maintain the pace when giants start copying their approach to swarm intelligence?